Search

This study focuses on the optimization process of crash box design. The design optimization process is resource‐intensive and requires multiple dynamic simulations. Numerous design parameters can be varied to satisfy the crashworthiness objectives. Therefore an intelligent and robust machine learning (ML) framework has been developed. This framework can be employed to assist in the optimization of various crashworthiness components with different crashworthiness objectives. Here, a reinforcement learning‐based (RL) machine learning framework is developed. It consists of a finite element method (FEM) surrogate and RL environment. The FEM surrogate is trained using data from FEM simulations as well as synthetic data generated by a Generative Adversarial Network (GAN). An inverse problem is solved for optimizing the geometrical parameters of the boundary value problem while the RL is receiving input from the FEM surrogate. RL has proven to be accurate in many fields due to its ability to explore and exploit the learned dynamics. Conventional algorithms require numerous iterations and tuned functions to achieve appropriate results and need to be initialized after a slight change in the problem. Due to an optimization of input parameters in an inverse problem, RL is chosen for the present investigation. For the simulation data needed, the crash box is designed with linear first‐order accurate four nodal shell elements. Also, bilinear elastoplastic material along with geometrical nonlinearity is used to simulate the dynamic deformation process. The RL agents learn to vary the crash box design parameters and search for the optimal parameters. The optimal parameters are based on the user‐defined crashworthiness objectives. [ABSTRACT FROM AUTHOR]

The primary goal of the current study is to optimise crash box designs for automobiles. Machine learning (ML) techniques are used to build an intelligent and reliable ML framework. With the help of this framework, the crash box design can be optimised for crashworthiness analysis. The optimisation of the crash box design is resource‐intensive due to its intricate geometric design, use of a variety of materials, and extensive use of dynamic simulations to determine the ideal structural parameters through simulations. A reinforcement learning‐based (RL) optimization technique is developed for this reason. The crash box is built with nonlinear first‐order shear deformation shell elements to provide the necessary simulation data, and nonlinear elastoplastic material is used to model the dynamic impact process. The RL agents are tuned using finite element (FE) simulations and synthetic data generated by a generative adversarial network (GAN) to optimise and select the optimal model parameters for vehicle crashworthiness analysis. To estimate the correct model parameter required to fulfil specified crashworthiness metrics, the RL agent learns and optimises automatically. This method is effective in terms of resources and could be helpful in the early stages of vehicle development. [ABSTRACT FROM AUTHOR]